Value Alignment
Embedding what we care about—fairness, honesty, harm reduction—into the model's behavior.
Think of it like
Like teaching a child your family's principles so they know right from wrong when you're not around.
Example
A hiring model is taught not to discriminate based on gender, even if historical data shows a pattern.
How it actually works
The challenge is that values aren't neutral or universal. What counts as 'honest' or 'fair' varies by culture and context. You can encode rules (don't discriminate on gender), but that's brittle. You can fine-tune on examples of good behavior, but that's expensive and incomplete. The best you can do is be explicit about whose values you're encoding and accept that you'll make tradeoffs.
For product teams
Define your product's values upfront and evaluate models against them, not just accuracy.
For engineers
Use instruction tuning, constitutional AI, and adversarial testing to enforce value-aligned behavior.
Related
- Alignment Research — Broader goal.
Read anything AI without the jargon
Look up any term in plain English, or save terms as you read with the free Chrome extension.
Open DecoderAdd to Chrome